I fine-tuned my dataset using the pi3. I adopted a high-resolution configuration and disabled the confidence head. I noticed that the pose quickly converged, but the points became blurry. However, when using the pre-trained weights of pi3 for prediction, the local point cloud for each frame had a normal shape. Why is that? What are the things to note during the fine-tuning process?
（This is the fintuned result.）
(This is the pretrained model‘s result.)
The pretraining results have a general shape (perhaps the depth is not precise enough), but the pose is incorrect. Now, after fine-tuning, the results have a more accurate pose, but the depth is lost. And the training loss even still lower, and the validation loss is also much lower than the first few training iterations but remains unchanged after few epochs.
I fine-tuned my dataset using the pi3. I adopted a high-resolution configuration and disabled the confidence head. I noticed that the pose quickly converged, but the points became blurry. However, when using the pre-trained weights of pi3 for prediction, the local point cloud for each frame had a normal shape. Why is that? What are the things to note during the fine-tuning process?
The pretraining results have a general shape (perhaps the depth is not precise enough), but the pose is incorrect. Now, after fine-tuning, the results have a more accurate pose, but the depth is lost. And the training loss even still lower, and the validation loss is also much lower than the first few training iterations but remains unchanged after few epochs.